[webgpu] add util functions for creating tensor view (microsoft#24566)

### Description

The added util functions can be used in 2 ways:

- create a reshaped tensor from an existing one.
- create a reinterpret view of a different type (will be useful in
(u)int4/(u)int8 operator implementation)

Author: fs-eire

onnxruntime/core/providers/webgpu/compute_context.h

@@ -88,6 +88,9 @@ class ComputeContext {
   //
   // Create CPU tensor.
   //
+  // This method creates a tensor of the given data type and shape, using the CPU allocator.
+  // The tensor owns the underlying CPU memory buffer.
+  //
   template <typename TensorShapeType>
   Tensor CreateCPUTensor(MLDataType data_type, TensorShapeType&& shape) {
     AllocatorPtr allocator;
@@ -98,6 +101,9 @@ class ComputeContext {
   //
   // Create GPU tensor.
   //
+  // This method creates a tensor of the given data type and shape, using the WebGPU allocator.
+  // The tensor owns the underlying WebGPU storage buffer.
+  //
   template <typename TensorShapeType>
   Tensor CreateGPUTensor(MLDataType data_type, TensorShapeType&& shape) {
     AllocatorPtr allocator;

onnxruntime/core/providers/webgpu/webgpu_utils.h

@@ -5,11 +5,15 @@
 
 #include <cstdint>
 #include "core/common/common.h"
+#include "core/framework/tensor.h"
 #include "core/framework/tensor_shape.h"
 
 namespace onnxruntime {
 namespace webgpu {
 
+/**
+ * Returns the maximum number of components `N` to be used as `vecN` for the given size.
+ */
 inline int GetMaxComponents(int64_t size) {
   if (size % 4 == 0) {
     return 4;
@@ -19,6 +23,11 @@ inline int GetMaxComponents(int64_t size) {
   return 1;
 }
 
+/**
+ * Returns a string representing a WGSL expression that sums the components of a value T.
+ *
+ * T can be a scalar S, vec2<S> or vec4<S>.
+ */
 inline std::string SumVector(std::string x, int components) {
   switch (components) {
     case 1:
@@ -49,5 +58,36 @@ inline std::string MakeScalarOrVectorType(int components, std::string_view data_
 
 TensorShape ReduceShapeByComponents(const TensorShape& shape, int64_t components);
 
+/**
+ * Create a reshaped tensor from an existing tensor.
+ *
+ * The specified new shape must have the same number of elements as the original tensor.
+ *
+ * The new tensor is a "view" of the original tensor. It uses the same data of the original tensor.
+ * The new tensor does not take or share ownership of the underlying data. The original tensor must outlive the new tensor.
+ */
+inline Tensor CreateTensorView(const Tensor& tensor, const TensorShape& new_shape) {
+  ORT_ENFORCE(tensor.Shape().Size() == new_shape.Size(), "Cannot reshape tensor ", tensor.Shape().ToString(), " to ", new_shape.ToString());
+  return {tensor.DataType(), new_shape, const_cast<void*>(tensor.DataRaw()), tensor.Location()};
+}
+
+/**
+ * Create a reinterpreted tensor from an existing tensor with a new data type and shape.
+ *
+ * The new data type and shape must match the original tensor's storage size.
+ *
+ * The new tensor is a "view" of the original tensor. It uses the same data of the original tensor.
+ * The new tensor does not take or share ownership of the underlying data. The original tensor must outlive the new tensor.
+ */
+inline Tensor CreateTensorView(const Tensor& tensor, MLDataType new_data_type, const TensorShape& new_shape) {
+  auto byte_size = Tensor::CalculateTensorStorageSize(tensor.DataType(), tensor.Shape());
+  auto new_byte_size = Tensor::CalculateTensorStorageSize(new_data_type, new_shape);
+  ORT_ENFORCE(byte_size == new_byte_size,
+              "Cannot reshape tensor ", tensor.Shape().ToString(), " to ", new_shape.ToString(),
+              " with data type ", DataTypeImpl::ToString(new_data_type), ". The byte size of the original tensor is ",
+              byte_size, " and the byte size of the new tensor is ", new_byte_size);
+  return {new_data_type, new_shape, const_cast<void*>(tensor.DataRaw()), tensor.Location()};
+}
+
 }  // namespace webgpu
 }  // namespace onnxruntime